A new proof-of-concept project called ACP P2P is attempting to solve one of the more interesting problems in the AI agent space: how do autonomous agents find and communicate with each other without relying on centralized infrastructure? Developer skotkiewicz dropped the implementation on Hacker News this week, describing it as a peer-to-peer take on something like context7—but provided natively by agents for agents.

The Core Concept Behind ACP P2P

The Agent Client Protocol (ACP) P2P implementation allows any two agents on the network to discover each other via mDNS and establish direct communication channels. Instead of routing requests through an API gateway or third-party service, agents can broadcast questions to all connected peers using commands like ask or target specific agents based on their declared capabilities with syntax such as ask --caps rust,math . The system uses libp2p's gossipsub protocol for message propagation across the mesh, meaning information spreads organically without a single point of failure.

Technical Architecture

The project is written entirely in Rust, leveraging Tokio for async runtime support and clap for command-line argument parsing. Each agent has a human-readable alias, can advertise specific capabilities via tags like rust, math, or python, and maintains its own cryptographic peer identity generated at startup (or ephemeral if the --ephemeral flag is used). The ACP feature enables JSON-RPC payload broadcasting between agents, making this potentially useful as infrastructure for more complex multi-agent workflows. Building requires running cargo build --release --features acp from the cloned repository.

Why This Matters for Agent-to-Agent Communication

Current AI agent ecosystems often rely on centralized orchestration layers—LangChain, AutoGen, or similar frameworks that coordinate agent behavior through a single control plane. ACP P2P represents an alternative paradigm where agents are genuinely peer-to-peer: they discover each other dynamically, advertise what they can do, and respond to requests from any compatible peer. This approach mirrors how BitTorrent disrupted centralized file sharing—though whether decentralized agent communication catches on remains to be seen given the project's modest reception (just 2 points on HN at publication time).

Limitations and Open Questions

The project is explicitly labeled as a proof-of-concept, so production readiness shouldn't be expected. There's no mention of authentication mechanisms, access control for sensitive queries, or quality assurance for responses from untrusted peers—legitimate concerns given that any agent can respond to broadcasts on the network. The zero visible comments suggest this is either very early-stage work or addressing a problem that hasn't gained widespread mindshare yet among developers.

Key Takeaways

  • Built with Rust + libp2p for robust P2P networking foundations
  • Capability-based routing lets agents target specific expertise without manual peer discovery
  • mDNS enables automatic local network peer detection out of the box
  • Gossipsub protocol handles message propagation across distributed agent mesh
  • MIT licensed and open to contributions on GitHub

The Bottom Line

ACP P2P is a hacker-built experiment in decentralized AI agent communication—technically solid but niche. Whether it gains traction depends entirely on whether the developer community sees value in cutting out the orchestration middleman. Worth watching, but don't bet your infrastructure on it just yet.